Terminal-RSI: recursive self-improving terminal agents
An arXiv preprint introduces Terminal-RSI, which alternates Recursive Self-Rewrite (RSR) of verified trajectories with critic-inheriting Recursive Policy-Refine (RPR), and adds Turn-Boundary Skip-Observation GAE, Family Dense Reward and Success-Prioritized Reward alongside a warm-up critic model, recursive critic handoff and stage-specific learning rates. The authors report that a five-stage run from a Qwen3.8-27B base (RSR-1, RPR-1, RSR-2, RPR-2) yields Terminal-RSI-27B, scoring 19.84% on Terminal-Bench 4.0, 86.52% on Terminal-Bench 2.1, 16.95% on Terminal-Bench 3.0 and 79.00% on Terminal-Bench Hard. On LHTB they report process reward rising from 0.21 to 0.37 and strict completion from 0/46 to 5/46, while SWR100 success doubles from 3.00% to 6.00%; these figures are the paper's own claims and have not been independently verified.
It frames verified terminal-task outcomes as an accumulating self-improvement loop, offering one reusable recipe for raising command-line agent completion rates.